Nama : Debbi Indriana Dewi (08011381320012)

Original Article
OR Insight (2009) 22, 201–220. doi:10.1057/ori.2009.10

Analytic Hierarchy Process and Expert Choice: Benefits and

limitations
Alessio Ishizaka1 and Ashraf Labib1
Portsmouth Business School, University of Portsmouth, Richmond Building, Portland Street, Portsmouth PO1 3DE, UK
1

Correspondence: Alessio Ishizaka, E-mails: [email protected]; [email protected]

Topof page
Abstract
This article describes the original Analytic Hierarchy Process (AHP) as it is implemented in the software
package Expert Choice. We demonstrate its application through a practical example. In particular, we
discuss problem modelling, pairwise comparisons, judgement scales, derivation methods, consistency
indices, synthesis of the weights and sensitivity analysis. Finally, the limitations of the original AHP along
with the new proposed development are explained.

Keywords:

AHP, decision making, review

Topof page
Introduction
The Analytic Hierarchy Process (AHP) is a multi-criteria decision making (MCDM) method that helps the
decision-maker facing a complex problem with multiple conflicting and subjective criteria (for example location
or investment selection, projects ranking and so forth). Several papers have compiled the AHP success stories in
very different fields (Zahedi, 1986; Golden et al, 1989; Shim, 1989; Vargas, 1990; Saaty and Forman,
1992; Forman and Gass, 2001; Kumar and Vaidya, 2006; Omkarprasad and Sushil, 2006; Ho, 2008; Liberatore
and Nydick, 2008). The oldest reference we have found dates from 1972 (Saaty, 1972). After this, a paper in
the Journal of Mathematical Psychology (Saaty, 1977) precisely described the method. The vast majority of the
applications still use AHP as described in this first publication and are unaware of successive developments.
This fact is probably owing to the leading software supporting AHP, namely, Expert Choice
(http://www.expertchoice.com/), which still incorporates AHP as it was described in its first publication. In this
article, we describe AHP through Expert Choice and provide a sketch of the major directions in methodological
developments (as opposed to a discussion of applications) and the further research in this important field.

Topof page
The Original AHP Method

Like several other MCDM methods such as ELECTRE, MacBeth, SMART, PROMETHEE, UTA and
so on (Belton and Stewart, 2002; Figueira et al, 2005), AHP is based on four steps: problem
modelling, weights valuation, weights aggregation and sensitivity analysis. In the next sections, we
will review these four steps used by AHP and its evolutions based on a simple problem: the selection
of a car to buy.

Problem modelling
As with all decision-making processes, the facilitator will sit a long time with the decision-maker(s) to
structure the problem, which can be divided into three parts: goal (buy a car), criteria (initial cost,
maintenance cost, prestige, quality and its sub-criteria) and alternatives (Fiat Uno, Nissan Maxima 4
Doors, Mercedes Benz 290, Volvo 840, Ford Fiesta) (Figure 1). AHP has the advantage of permitting
a hierarchical structure of the criteria, which provides users with a better focus on specific criteria and
sub-criteria when allocating the weights.
Figure 1.
 Example of hierarchy.

Pairwise comparisons
At each node of the hierarchy, a matrix will collect the pairwise comparisons of the decision-maker (for
example Figure 2).
Figure 2.

Comparison matrix of the first node.

Psychologists argue that it is easier and more accurate to express one's opinion on only two alternatives than
simultaneously on all the alternatives. It also allows consistency and cross checking between the different
pairwise comparisons (see Consistency part under the section The original AHP method). AHP uses a ratio
scale, which, contrary to methods using interval scales (Kainulainen et al, 2009), requires no units in the
comparison. The judgement is a relative value or a quotient a/b of two quantities a and b having the same units
(intensity, meters, utility and so on). The decision-maker does not need to provide a numerical judgement;
instead a relative verbal appreciation, more familiar in our daily lives, is sufficient.

If the matrix is perfectly consistent, then the transitivity rule (1) holds for all comparisons aij:

For example, if team A beats team B two-zero and team B beats team C three-zero, then it is expected with the
transitivity rule (1) that team A beats team C six-zero (3·2=6). However, this is seldom the case because our
world is inconsistent by nature. As a minimal consistency is required to derive meaningful priorities, a test must
be done (see Consistency part under the section The original AHP method). Webber et al (1996) state that the
order in which the comparisons are entered in the matrix may affect the successive judgments.

Judgement scales
One of AHP's strengths is the possibility to evaluate quantitative as well as qualitative criteria and alternatives
on the same preference scale of nine levels. These can be numerical (Figure 3), verbal (Figure 4) or graphical
(Figure 5).
Figure 3.

Numerical scale.

Figure 4.

Verbal scale.
Figure 5.

Graphical scale.

Priorities derivation
Once the comparisons matrices are filled, priorities can be calculated. The traditional AHP uses the eigenvalue
method. For some users this method seems quite obscure. In order to explain it, we start from the case of a
consistent matrix with known priorities pi. In this case, the comparison of the alternatives i and j is given by pi/pj,
which multiplied by the priority vector results in:

or grouped:

where vector of the priorities; n: dimension of the matrix; A: comparison matrix

Equation (2) is the formulation of an eigenvector problem. The calculated priorities are exact for a consistent
matrix. When slight inconsistencies are introduced, priorities should vary only slightly according to the
perturbation theory (Saaty, 2003).

Consistency
As priorities make sense only if derived from consistent or near consistent matrices, a consistency check must
be applied. Saaty (1977) has proposed a consistency index (CI), which is related to the eigenvalue method:

where λmax=maximal eigenvalue

The consistency ratio, the ratio of CI and RI, is given by:

where RI is the random index (the average CI of 500 randomly filled matrices).

If CR is less than 10 per cent, then the matrix can be considered as having an acceptable consistency.

Saaty (1977) calculated the RIs shown in Table 1.

Table 1 - Random indices from Saaty (1977).

Other researchers have run simulations with different numbers of matrices (Lane and Verdini, 1989; Tummala
and Wan, 1994; Alonso and Lamata, 2006) or incomplete matrices (Forman, 1990). Their RIs are different but
close to Saaty's.
Aggregation
The last step is to synthesize the local priorities across all criteria in order to determine the global priority. The
historical AHP approach (called later distributive mode) adopts an additive aggregation (5) with normalization
of the sum of the local priorities to unity (Figure 6):
Figure 6.

Priorities with the distributive mode.

where pi: global priority of the alternative i; lij: local priority; wj: weight of the criterion j.

The ideal mode uses a normalization by dividing the score of each alternative only by the score of the best
alternative under each criterion (Figure 7).

Figure 7.

Priorities with the ideal mode.

It can be seen that the two approaches do not necessarily recommend the same ranking (see the rectangles
in Figures 6 and 7).

If we already know the priorities, the distributive mode is the only approach that will retrieve these priorities.
However, if a copy (Belton and Gear, 1983) or a near copy (Dyer, 1990b) of an alternative is introduce or
removed (Troutt, 1988), a rank reversal of the alternatives may appear. This phenomenon has been criticized
from one side (Dyer, 1990a, 1990b; Holder, 1990, 1991; Stam and Duarte Silva, 2003) and legitimized by
another (Saaty, 1986,1990, 1991, 1994, 1995, 2006; Harker and Vargas, 1987, 1990; Pérez, 1995). This rank
reversal phenomenon is not unique to AHP but to all additive models (Triantaphyllou, 2001; Wang and Luo,
2009).

Millet and Saaty (2000) gave some guidance on which normalization to use. If we are in a closed system (that is
no alternative will be added or removed), then the distributive mode should be used. If we are in an open system
(that is alternatives can be added or removed) and we allow our preferences for alternatives to be dependent on
other alternatives (in other words, we accept the rank reversal phenomenon), then the distributive mode is
indicated. If we are in an open system and you do not want that other alternatives to affect the outcome, then the
ideal mode is recommended. Based on these observations, Expert Choice has been modified to support the two
modes.

Sensitivity analysis
The last step of the decision process is the sensitivity analysis, where the input data are slightly modified in
order to observe the impact on the results. If the ranking does not change, the results are said to be robust. The
sensitivity analysis is best performed with an interactive graphical interface. Expert Choice allows different
sensitivity analyses, where the main difference is the various graphical representations (Figure 8).
Figure 8.

An example of four possible graphical sensitivity analyses in Expert Choice.

Topof page
Limitations of the Original AHP and Expert Choice

Problem structuring
This step is important, because a different structure may lead to a different final ranking. Several authors
(Stillwell et al, 1987; Weber et al, 1988; Pöyhönen et al, 1997) have observed that criteria with a large number
of sub-criteria tend to receive more weight than when they are less detailed. Brugha (2004) has provided a
complete guideline to structure a problem hierarchically. A book (Saaty and Forman, 1992) compiling
hierarchies in different applications has been written. When setting up the AHP hierarchy with a large number
of elements, the decision-maker should attempt to arrange these elements in clusters so they do not differ in
extreme ways (Saaty, 1991; Ishizaka, 2004a,2004b).

Pairwise comparisons
Comparisons are recorded in a positive reciprocal matrix, as in Figure 2. In special cases, such as in currencies
exchanges, not reciprocal matrices can be used (Hovanov et al, 2008). Even if not reciprocal matrices are then
treated similarly to traditional matrices, Expert Choice does not offer the possibility not to be reciprocal.

Judgement scales
The use of verbal comparisons is intuitively appealing, user-friendly and more common in our everyday lives
than numbers. It may also allow some ambiguity in non-trivial comparisons. This ambiguity in the English
language has also been criticized (Donegan et al, 1992). Owing to its pairwise comparisons AHP needs ratio
scales. Barzilai (2005) claims that preferences cannot be represented with ratio scales, because in his opinion an
absolute zero does not exists, as with temperature or electrical tension. Saaty (1994) states that ratio scales are
the only possible measurement if we want to be able to aggregate measurement, as in a weighted sum. Dodd and
Donegan (1995) have criticized the absence of a zero in the preference scale.

To derive priorities, the verbal comparisons must be converted into numerical ones. In Saaty's AHP the verbal
statements are converted into integers from one to nine. Theoretically there is no reason to be restricted to these
numbers and verbal gradation. If the verbal gradation has been little investigated, several other numerical scales
have been proposed (Table 2; Figures 9 and 10). Harker and Vargas (1987) have evaluated a quadratic and a
root square scale in only one simple example and argued in favour of Saaty's 1–9 scale. However, one example
seems not enough to conclude the superiority of the 1–9 linear scale. Lootsma (1989) argued that the geometric
scale is preferable to the 1–9 linear scale. Salo and Hämäläinen (1997) point out that the integers from one to
nine yield local weights, which are unevenly dispersed, so that there is lack of sensitivity when comparing
elements, which are preferentially close to each other. Based on this observation, they propose a balanced scale
where the local weights are evenly dispersed over the weight range [0.1, 0.9]. Earlier Ma and Zheng (1991) have
calculated a scale where the inverse elements x of the scale 1/x are linear instead of the x in the Saaty
scale. Donegan et al (1992) have proposed an asymptotic scale avoiding the boundary problem, for example if
the decision-maker enters aij=3 and ajk=4, she/he is forced to an intransitive relation (1) because the upper limit
of the scale is 9 and she/he cannot enter aik=12. Ji and Jiang (2003) propose a mixture of verbal and geometric
scales. The possibility to integrate negative values in the scale has been also explored (Saaty and Ozdemir,
2003; Millet and Schoner, 2005).
Figure 9.

Graph of the judgement scales.

Figure 10.

Graph of the judgement scales without the geometric and power scales.

Table 2 - Different scales for comparing two alternatives.

Among all the proposed scales, the linear scale with the integers one to nine and their reciprocals has been used
by far the most often in applications. It is also the only one implemented in Expert Choice. Saaty (1980, 1991)
advocates it as the best scale to represent weight ratios. However, the cited examples deal with objective
measurable alternatives such as the areas of figures, whereas AHP mainly treats decision processes as subjective
issues. We understand the difficulty of verifying the effectiveness of scales through subjective issues. Salo and
Hämäläinen (1997) demonstrate the superiority of the balanced scale when comparing two elements. The choice
of the ‘best’ scale is a very heated debate. Some scientists argue that the choice depends on the person and the
decision problem (Harker and Vargas, 1987; Pöyhönen et al, 1997). Therefore, we believe that other scales
would be welcomed in Expert Choice.

Priorities derivation
Only 2 years later than the publication of the original AHP, Johnson et al (1979) show a rank reversal problem
for scale inversion with the eigenvalue method. For example, consider the matrix of Figure 11, asking which car
is the most fuel economical. The calculated priorities are: (0.061; 0.374; 0.134; 0.387; 0.043), with Volvo 840
being the winner.
Figure 11.

Comparison matrix.

If we invert the question: which car has the highest fuel consumption, then the comparisons are simply inverted
(Figure 12). The calculated priorities are: (0.299; 0.047; 0.140; 0.051; 0.462), with Nissan being the winner. We
have a rank reversal because of the formulation of the problem (which is a different and independent cause of
the rank reversal because of the introduction or deletion of an alternative).

Figure 12.
 Comparisons of Figure 11 are inverted.

In order to avoid this problem, Crawford and Williams (1985) have adopted another approach in minimizing the
multiplicative error (3):

where aij is the comparison between object i and j; pi is the priority of object i; eij is the error

The multiplicative error is commonly accepted to be log normal distributed (similarly the additive error would
be assumed to be normal distributed). The geometric mean (7) will minimize the sum of these errors (8).

The geometric mean (also sometimes known as Logarithmic Least Squares Method) can be easily calculated by
hand and has been supported by a large segment of the AHP community (Budescu, 1984; Fichtner,
1986; Lootsma, 1993,1996; Barzilai, 1997; Barzilai and Lootsma, 1997; Aguarón and Moreno-Jiménez,
2000; Escobar and Moreno-Jiménez, 2000; Aguarón and Moreno-Jiménez, 2003; Leskinen and Kangas, 2005).
Its main advantage is the absence of rank reversals because of the right and left inconsistency, In fact geometric
means of rows and columns provide the same ranking (which is not necessarily the case with the eigenvalue
method).

If mathematical evidence points clearly to the geometric mean over the eigenvalue method, there is no clear
differences between these two methods when simulations are applied (Budescu et al, 1986; Golany and Kress,
1993;Herman and Koczkodaj, 1996; Cho and Wedley, 2004; Jones and Mardle, 2004; Ishizaka and Lusti, 2006),
apart from special cases (Bajwa et al, 2008). Perhaps in the light of this lack of practical evidence, Saaty's group
has always supported the eigenvalue method (Saaty and Vargas, 1984a, 1984b; Harker and Vargas, 1987; Saaty
and Hu, 1998;Saaty, 2003).

Other methods have been proposed, each one based either on the idea of the distance minimization (like the
geometric mean) or on the idea that small perturbation inducing small errors (like the eigenvalue method or the
arithmetic mean of rows). Cho and Wedley (2004) have enumerated 18 different methods, which are effectively
15 because three are equivalent to others (Lin, 2007). However, Expert Choice incorporates only the eigenvalue
method.

Consistency
Expert Choice uses the consistency ratio (4). However, this consistency ratio has been criticized because it
allows contradictory judgements in matrices (Kwiesielewicz and van Uden, 2004; Bana e Costa and Vansnick,
2008) or rejects reasonable matrices (Karapetrovic and Rosenbloom, 1999). Interactive techniques based on the
transitivity rule (1) have been developed in order to discover contradictory judgements and correct them
(Ishizaka and Lusti, 2004; Wanget al, 2009). Their inclusion in Expert choice would be beneficial.

Several other methods have been proposed to measure consistency. Peláez and Lamata (2003) describe a
method based on the determinant of the matrix. Crawford and Williams (1985) prefer to sum the difference
between the ratio of the calculated priorities and the given comparisons. The transitivity rule (2) has been used
by Salo and Hämäläinen (1997) and later by Ji and Jiang (2003). Alonso and Lamata (2006) have computed a
regression of the RIs and propose the formulation:

Stein and Mizzi (2007) use the normalized column of the comparison matrix.

For all consistency checking, some questions remain: what is the cut-off rule to declare my matrix inconsistent?
Should this rule depend on the size of the matrix? How should I adapt my consistency definition, when I use
another judgement scale?

Aggregation
The multiplicative aggregation (10) has been proposed to prevent the rank reversal phenomenon observed in the
distributive mode (see Aggregation part under the section The original AHP method) (Lootsma, 1993; Barzilai
and Lootsma, 1997).

The multiplicative aggregation has non-linearity properties allowing a superior compromise to be selected,
which is not the case with the additive aggregation (Stam and Duarte Silva, 2003; Ishizaka et al, 2006).
However, Vargas (1997)showed that additive aggregation is the only way to retrieve exact weights of known
objects. A multiplicative aggregation is not possible at the moment in Expert Choice.

Sensitivity analysis
The sensitivity analysis in Expert Choice varies the weights of the criteria as input data. It is also imaginable to
have in future a sensitivity analysis by varying interactively the local priorities of the alternatives (there is no
mathematical challenge in it).

However, sensitivity analysis is a fundamental process in the decision with AHP; it has received little attention
from the academic literature.

Topof page
Conclusion and Future Developments
Decisions that need support methods are difficult by definition and therefore complex to model. A trade-off
between prefect modelling and usability of the model should be achieved. It is our belief that AHP has reached
this compromise and will be useful for many other cases as it has been in the past. In particular, AHP has broken
through the academic community to be widely used by practitioners. This widespread use is certainly owing to
its ease of applicability and the structure of AHP, which follows the intuitive way in which managers solve
problems. The hierarchical modelling of the problem, the possibility to adopt verbal judgements and the
verification of the consistency are its major assets. Expert Choice, the user-friendly supporting software, has
certainly largely contributed to the success of the method. It incorporates intuitive graphical user interfaces,
automatic calculation of priorities and inconsistencies and several ways to process a sensitivity analysis. Today,
several other supporting software packages have been developed: Decision Lens, HIPRE 3+, RightChoiceDSS,
Criterium, EasyMind, WebAHP, not to mention that a template in Excel could also be easily generated.
However, most of the software are based only on the classical AHP. Along with its traditional applications, a
new trend, as compiled by the work of Ho (2008), is to use AHP in conjunction with others methods:
mathematical programming techniques like linear programming, Data Envelopment Analysis, Fuzzy Sets,
House of Quality, Genetic Algorithms, Neural Networks, SWOT-analysis, and recently with GIS (Temiz and
Tecim, 2009). There is little doubt that AHP will be more and more frequently adopted.

AHP still suffers from some theoretical disputes. The rank reversal is surely the most debated problem. This
phenomenon is still not fully resolved and maybe it will never be because the aggregation of preferences
transposed from scales of different units is not easily interpretable and even questionable according to the
French school (Roy, 1996).

The assumption of criteria independence (no correlation) may be sometimes a limitation of AHP (and other
MCDM methods). The Analytic Network Process (ANP), a generalization of AHP with feedbacks to adjust
weights, may be a solution. However, the decision-maker must answer a much larger number of questions,
which may be quite complex: for example ‘Given an alternative and a criterion, which of the two alternatives
influences the given criterion more and how much more than another alternative’ (Saaty and Takizawa, 1986).
A simplified ANP, while still keeping its proprieties, would be beneficial for a wider adoption of the method.
Another direction of the research will probably be on a more soft side. The choice of a hierarchy and a
judgement scale is important and difficult. Problem structuring methods could help in the construction of AHP
hierarchies.